1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8
9 #include <assert.h>
10
11 #include <immintrin.h>
12
13 #include <xnnpack/gemm.h>
14
15
xnn_q8_gemm_ukernel_4x4c2__sse2(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_q8_gemm_params params[restrict static1])16 void xnn_q8_gemm_ukernel_4x4c2__sse2(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const uint8_t* restrict a,
21 size_t a_stride,
22 const void* restrict w,
23 uint8_t* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_q8_gemm_params params[restrict static 1])
27 {
28 assert(mr != 0);
29 assert(mr <= 4);
30 assert(nc != 0);
31 assert(kc != 0);
32
33 const uint8_t* a0 = a;
34 uint8_t* c0 = c;
35 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
36 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
37 if XNN_UNPREDICTABLE(mr < 2) {
38 a1 = a0;
39 c1 = c0;
40 }
41 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
42 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
43 if XNN_UNPREDICTABLE(mr <= 2) {
44 a2 = a1;
45 c2 = c1;
46 }
47 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
48 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
49 if (mr != 4) {
50 a3 = a2;
51 c3 = c2;
52 }
53
54 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->sse2.kernel_zero_point);
55
56 do {
57 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
58 __m128i vacc1x0123 = vacc0x0123;
59 __m128i vacc2x0123 = vacc0x0123;
60 __m128i vacc3x0123 = vacc0x0123;
61 w = (const void*) ((uintptr_t) w + 16);
62
63 const __m128i vzero = _mm_setzero_si128();
64 size_t k = kc;
65 while (k >= 8 * sizeof(uint8_t)) {
66 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
67 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
68 a0 += 8;
69 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
70 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
71 a1 += 8;
72 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
73 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
74 a2 += 8;
75 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
76 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
77 a3 += 8;
78
79 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
80 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
81
82 vacc0x0123 = _mm_add_epi32(vacc0x0123,
83 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
84 vacc1x0123 = _mm_add_epi32(vacc1x0123,
85 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
86 vacc2x0123 = _mm_add_epi32(vacc2x0123,
87 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
88 vacc3x0123 = _mm_add_epi32(vacc3x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
90
91 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
92 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
96 vacc1x0123 = _mm_add_epi32(vacc1x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
98 vacc2x0123 = _mm_add_epi32(vacc2x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
100 vacc3x0123 = _mm_add_epi32(vacc3x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
102
103 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
104 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
105
106 vacc0x0123 = _mm_add_epi32(vacc0x0123,
107 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
108 vacc1x0123 = _mm_add_epi32(vacc1x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
110 vacc2x0123 = _mm_add_epi32(vacc2x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
112 vacc3x0123 = _mm_add_epi32(vacc3x0123,
113 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
114
115 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
116 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
117 w = (const void*) ((uintptr_t) w + 32);
118
119 vacc0x0123 = _mm_add_epi32(vacc0x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
121 vacc1x0123 = _mm_add_epi32(vacc1x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
123 vacc2x0123 = _mm_add_epi32(vacc2x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
125 vacc3x0123 = _mm_add_epi32(vacc3x0123,
126 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
127
128 k -= 8 * sizeof(uint8_t);
129 }
130 if (k != 0) {
131 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
132 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
133 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
134 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
135 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
136 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
137 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
138 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
139 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
140 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
141 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
142 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
143
144 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
145 w = (const void*) ((uintptr_t) w + 8);
146 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
147
148 vacc0x0123 = _mm_add_epi32(vacc0x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
150 vacc1x0123 = _mm_add_epi32(vacc1x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
152 vacc2x0123 = _mm_add_epi32(vacc2x0123,
153 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
154 vacc3x0123 = _mm_add_epi32(vacc3x0123,
155 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
156
157 if (k > 2 * sizeof(uint8_t)) {
158 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
159 w = (const void*) ((uintptr_t) w + 8);
160 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
161
162 vacc0x0123 = _mm_add_epi32(vacc0x0123,
163 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
164 vacc1x0123 = _mm_add_epi32(vacc1x0123,
165 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
166 vacc2x0123 = _mm_add_epi32(vacc2x0123,
167 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
168 vacc3x0123 = _mm_add_epi32(vacc3x0123,
169 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
170
171 if (k > 4 * sizeof(uint8_t)) {
172 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
173 w = (const void*) ((uintptr_t) w + 8);
174 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
175
176 vacc0x0123 = _mm_add_epi32(vacc0x0123,
177 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
178 vacc1x0123 = _mm_add_epi32(vacc1x0123,
179 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
180 vacc2x0123 = _mm_add_epi32(vacc2x0123,
181 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
182 vacc3x0123 = _mm_add_epi32(vacc3x0123,
183 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
184
185 if (k > 6 * sizeof(uint8_t)) {
186 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) w);
187 w = (const void*) ((uintptr_t) w + 8);
188 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
189
190 vacc0x0123 = _mm_add_epi32(vacc0x0123,
191 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
192 vacc1x0123 = _mm_add_epi32(vacc1x0123,
193 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
194 vacc2x0123 = _mm_add_epi32(vacc2x0123,
195 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
196 vacc3x0123 = _mm_add_epi32(vacc3x0123,
197 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
198 }
199 }
200 }
201 }
202
203 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
204 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
205
206 const __m128i vnmask0x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0x0123);
207 const __m128i vnmask1x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc1x0123);
208 const __m128i vnmask2x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc2x0123);
209 const __m128i vnmask3x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc3x0123);
210
211 const __m128i vabsacc0x0123 = _mm_sub_epi32(_mm_xor_si128(vacc0x0123, vnmask0x0123), vnmask0x0123);
212 const __m128i vabsacc1x0123 = _mm_sub_epi32(_mm_xor_si128(vacc1x0123, vnmask1x0123), vnmask1x0123);
213 const __m128i vabsacc2x0123 = _mm_sub_epi32(_mm_xor_si128(vacc2x0123, vnmask2x0123), vnmask2x0123);
214 const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123);
215
216 const __m128i vabsacc0x1032 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(2, 3, 0, 1));
217 const __m128i vabsacc1x1032 = _mm_shuffle_epi32(vabsacc1x0123, _MM_SHUFFLE(2, 3, 0, 1));
218 const __m128i vabsacc2x1032 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(2, 3, 0, 1));
219 const __m128i vabsacc3x1032 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(2, 3, 0, 1));
220
221 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
222 const __m128i vabsprod1x02 = _mm_mul_epu32(vabsacc1x0123, vmultiplier);
223 const __m128i vabsprod2x02 = _mm_mul_epu32(vabsacc2x0123, vmultiplier);
224 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier);
225
226 const __m128i vnmask0x02 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(2, 2, 0, 0));
227 const __m128i vnmask1x02 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(2, 2, 0, 0));
228 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0));
229 const __m128i vnmask3x02 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(2, 2, 0, 0));
230
231 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02);
232 const __m128i vprod1x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x02, vnmask1x02), vnmask1x02);
233 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02);
234 const __m128i vprod3x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod3x02, vnmask3x02), vnmask3x02);
235
236 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
237 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31);
238 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31);
239 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31);
240
241 const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1032, vmultiplier);
242 const __m128i vabsprod1x13 = _mm_mul_epu32(vabsacc1x1032, vmultiplier);
243 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1032, vmultiplier);
244 const __m128i vabsprod3x13 = _mm_mul_epu32(vabsacc3x1032, vmultiplier);
245
246 const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));
247 const __m128i vnmask1x13 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(3, 3, 1, 1));
248 const __m128i vnmask2x13 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(3, 3, 1, 1));
249 const __m128i vnmask3x13 = _mm_shuffle_epi32(vnmask3x0123, _MM_SHUFFLE(3, 3, 1, 1));
250
251 const __m128i vprod0x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x13, vnmask0x13), vnmask0x13);
252 const __m128i vprod1x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x13, vnmask1x13), vnmask1x13);
253 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13);
254 const __m128i vprod3x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod3x13, vnmask3x13), vnmask3x13);
255
256 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31);
257 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31);
258 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31);
259 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31);
260
261 const __m128i vq31prod0x0213 = _mm_castps_si128(_mm_shuffle_ps(
262 _mm_castsi128_ps(vq31prod0x02), _mm_castsi128_ps(vq31prod0x13), _MM_SHUFFLE(2, 0, 2, 0)));
263 const __m128i vq31prod1x0213 = _mm_castps_si128(_mm_shuffle_ps(
264 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0)));
265 const __m128i vq31prod2x0213 = _mm_castps_si128(_mm_shuffle_ps(
266 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0)));
267 const __m128i vq31prod3x0213 = _mm_castps_si128(_mm_shuffle_ps(
268 _mm_castsi128_ps(vq31prod3x02), _mm_castsi128_ps(vq31prod3x13), _MM_SHUFFLE(2, 0, 2, 0)));
269
270 const __m128i vq31prod0x0123 = _mm_shuffle_epi32(vq31prod0x0213, _MM_SHUFFLE(3, 1, 2, 0));
271 const __m128i vq31prod1x0123 = _mm_shuffle_epi32(vq31prod1x0213, _MM_SHUFFLE(3, 1, 2, 0));
272 const __m128i vq31prod2x0123 = _mm_shuffle_epi32(vq31prod2x0213, _MM_SHUFFLE(3, 1, 2, 0));
273 const __m128i vq31prod3x0123 = _mm_shuffle_epi32(vq31prod3x0213, _MM_SHUFFLE(3, 1, 2, 0));
274
275 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
276
277 const __m128i vrem0x0123 =
278 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
279 const __m128i vrem1x0123 =
280 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
281 const __m128i vrem2x0123 =
282 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
283 const __m128i vrem3x0123 =
284 _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
285
286 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
287 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
288
289 vacc0x0123 =
290 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
291 vacc1x0123 =
292 _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
293 vacc2x0123 =
294 _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
295 vacc3x0123 =
296 _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
297
298 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
299 const __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
300 const __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
301 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
302 vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));
303 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));
304
305 if (nc >= 4) {
306 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
307 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_epi64(vout, 32));
308 *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(_mm_unpackhi_epi32(vout, vout));
309 *((uint32_t*) c3) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_si128(vout, 12));
310
311 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
312 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
313 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
314 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
315
316 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
317 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
318 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
319 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
320
321 nc -= 4;
322 } else {
323 if (nc & 2) {
324 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
325 c0 += 2;
326 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
327 c1 += 2;
328 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
329 c2 += 2;
330 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
331 c3 += 2;
332 vout = _mm_srli_epi32(vout, 16);
333 }
334 if (nc & 1) {
335 *((uint8_t*) c0) = (uint8_t) _mm_cvtsi128_si32(vout);
336 *((uint8_t*) c1) = (uint8_t) _mm_extract_epi16(vout, 2);
337 *((uint8_t*) c2) = (uint8_t) _mm_extract_epi16(vout, 4);
338 *((uint8_t*) c3) = (uint8_t) _mm_extract_epi16(vout, 6);
339 }
340
341 nc = 0;
342 }
343 } while (nc != 0);
344 }
345